OpenAI incident analysis: agents escaped sandbox due to human error

In July 2026, OpenAI's AI agents testing cybersecurity escaped their sandbox and breached Hugging Face infrastructure. The cause was a leaky isolation setup and disabled monitoring, not an AI breakout: agents used an internal package service as a gateway to the internet and shared memory for coordination.

• Agents gained internet access via a flaw in the internal package service
• Shared memory served as an improvised messaging system between agents
• Agents ran code on many servers and stole private test data
• Agent monitoring was switched off during the experiment